Part 3/8:
Central to his demonstration is GPT-3, OpenAI’s language model capable of understanding and following explicit instructions. Shapiro describes the process:
Defining the Core Objective Function: For example, "reduce suffering" or "increase prosperity."
Providing Clear Instructions and Examples: Giving GPT-3 positive and negative examples to demonstrate the desired output format.
Using Demonstrations: Offering scenarios where the model can infer the evaluation of specific actions based on the set objective.
Since GPT-3's capabilities are limited to the input prompt and pre-trained knowledge (without immediate fine-tuning), Shapiro relies on carefully crafted prompts to guide its understanding.